add fleet pure_fp16 into train.py (#5268)

PaddleNLP/benchmark/transformer/static/train.py

@@ -39,10 +39,9 @@ def do_train(args):
     paddle.enable_static()
     if args.is_distributed:
         fleet.init(is_collective=True)
-        gpu_id = int(os.getenv("FLAGS_selected_gpus", "0"))
-        places = paddle.CUDAPlace(
-            gpu_id) if args.use_gpu else paddle.static.cpu_places()
-        trainer_count = 1 if args.use_gpu else len(places)
+        places = [paddle.set_device("gpu")] if \
+                 args.use_gpu else paddle.static.cpu_places()
+        trainer_count = len(places)
     else:
         if args.use_gpu:
             places = paddle.static.cuda_places()
@@ -111,8 +110,10 @@ def do_train(args):
             if args.use_amp:
                 dist_strategy.amp = True
                 dist_strategy.amp_configs = {
-                    'custom_white_list': ['softmax', 'layer_norm', 'gelu'],
+                    'custom_white_list': ['softmax', 'layer_norm'],
                     'init_loss_scaling': args.scale_loss,
+                    'custom_black_list': ['lookup_table_v2'],
+                    'use_pure_fp16': args.use_pure_fp16
                 }
 
             optimizer = fleet.distributed_optimizer(
@@ -131,7 +132,7 @@ def do_train(args):
         optimizer.minimize(avg_cost)
 
     if args.is_distributed:
-        exe = paddle.static.Executor(places)
+        exe = paddle.static.Executor(places[0])
     else:
         exe = paddle.static.Executor()
         build_strategy = paddle.static.BuildStrategy()
@@ -144,9 +145,9 @@ def do_train(args):
                 exec_strategy=exec_strategy)
     exe.run(startup_program)
 
-    if not args.is_distributed and args.use_amp:
+    if args.use_amp:
         optimizer.amp_init(places[0])
-
+    
     # the best cross-entropy value with label smoothing
     loss_normalizer = -(
         (1. - args.label_smooth_eps) * np.log(